Predicting Guest Satisfaction for Airbnbs

STAT 301 – Group 41, Airbnb Prices in London, Rome & Budapest

Group Members:

• Rapeewit Chanprakaisi
• Renata Lovette
• Angelyca Purewal
• Yingqian Shi

Introduction

In recent years, the sharing economy has revolutionized the hospitality industry, with platforms like Airbnb offering travellers diverse and flexible lodging options across the globe. As Airbnb listings continue to grow, understanding the factors influencing guest satisfaction has become increasingly important for both hosts aiming to improve their offerings and for travellers seeking the best possible experience.

This project focuses on Airbnb listings across major European cities such as London, Rome, and Budapest. The dataset provides detailed information on various listing attributes, including price, location characteristics, listing type, host status, and more. In particular, it includes the guest_satisfaction_overall score, which serves as a summary measure of how satisfied guests were with their stay. Given the richness of this dataset, we aim to explore whether it is possible to predict overall guest satisfaction based on these various features.

While prior literature, such as the study by Zhu et al. (2020) and Rezazadeh Kalehbasti et al. (2021), has largely focused on predicting Airbnb prices using spatial and textual features, our work differs by shifting the outcome of interest to guest satisfaction. Their analysis demonstrated that machine learning models like XGBoost and SVR were effective for predicting prices, suggesting that similar approaches may also perform well in predicting satisfaction, though the underlying drivers may differ.

Question

We want to develop a model to figure out if we can predict guest satisfaction based on factors such as listing price, city (e.g., London, Rome, Budapest), location characteristics (e.g., latitude/longitude, distance from the city center), and Airbnb listing type (e.g., room type, business listing, multiple rooms or not).

Response Variable: guest_satisfaction_overall

Primary focus: Prediction – we are interested in understanding the relationship between all the variables and how they relate to and cause Airbnb guest satisfaction levels.

Data Description

The dataset includes different determinants of Airbnb prices across different European cities (London, Rome and Budapest) during the weekdays and weekends. For each city, the dataset includes 20 variables that were determined via spatial econometric analysis methods to analyse and identify the determinants of Airbnb prices across these cities.

Variable Name	Data Type	Description
...1	numerical - discrete	observation ID
realSum	numerical - continuous	price in Euros for a 2 night stay for 2 people
room_type	categorical	type of room (ie. shared, private)
room_shared	boolean	is the room shared?
room_private	boolean	is the room private?
person_capacity	numerical - discrete	maximum guest capacity in the room
host_is_superhost	boolean	does the host have superhost status?
multi	binary	does the property have multiple listings?
biz	binary	is the property hosted for business purposes?
cleanliness_rating	numerical - continuous	cleanliness rating of the listing
guest_satisfaction_overall	numerical - continuous	overall guest satisfaction rating
bedrooms	numerical - discrete	number of bedrooms
dist	numerical - continuous	distance from the city centre in km
metro_dist	numerical - continuous	distance from the nearest metro station in km
attr_index	numerical - continuous	attraction index of airbnb location
attr_index_norm	numerical - continuous	normalised attraction index of airbnb location (0-100)
rest_index	numerical - continuous	restaurant index of airbnb location
rest_index_norm	numerical - continuous	normalised restaurant index of airbnb location (0-100)
lng	numerical - continuous	longitude of airbnb location
lat	numerical - continuous	latitude of airbnb location

The number of observations for Airbnbs in...

• London are 4613 (weekdays) and 5378 (weekends)
• Rome are 4491 (weekdays) and 4534 (weekends)
• Budapest are 2073 (weekdays) and 1947 (weekends)

The data source and citation as requested by the owner(s) is

Gyódi, K., & Nawaro, Ł. (2021). Determinants of Airbnb prices in European cities: A spatial econometrics approach (Supplementary Material) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.4446043

Methods and Results

library(tidyverse)
library(repr)
library(ggplot2)
library(patchwork)
library(caret)
library(glmnet)
library(rsample)
library(car)
options(repr.plot.width = 16, repr.plot.height = 10)

Exploratory Data Analysis (EDA)

Read data

# Loading in the 6 data sets 
budapest_weekdays <- read_csv("https://raw.githubusercontent.com/rchanpra/stat-301-project/refs/heads/main/data/budapest_weekdays.csv")
budapest_weekends <- read_csv("https://raw.githubusercontent.com/rchanpra/stat-301-project/refs/heads/main/data/budapest_weekends.csv")
london_weekdays <- read_csv("https://raw.githubusercontent.com/rchanpra/stat-301-project/refs/heads/main/data/london_weekdays.csv")
london_weekends <- read_csv("https://raw.githubusercontent.com/rchanpra/stat-301-project/refs/heads/main/data/london_weekends.csv")
rome_weekdays <- read_csv("https://raw.githubusercontent.com/rchanpra/stat-301-project/refs/heads/main/data/rome_weekdays.csv")
rome_weekends <- read_csv("https://raw.githubusercontent.com/rchanpra/stat-301-project/refs/heads/main/data/rome_weekends.csv")

New names:
• `` -> `...1`
Rows: 2074 Columns: 20
── Column specification ────────────────────────────────────────────────────────
Delimiter: ","
chr  (1): room_type
dbl (16): ...1, realSum, person_capacity, multi, biz, cleanliness_rating, gu...
lgl  (3): room_shared, room_private, host_is_superhost

ℹ Use `spec()` to retrieve the full column specification for this data.
ℹ Specify the column types or set `show_col_types = FALSE` to quiet this message.
New names:
• `` -> `...1`
Rows: 1948 Columns: 20
── Column specification ────────────────────────────────────────────────────────
Delimiter: ","
chr  (1): room_type
dbl (16): ...1, realSum, person_capacity, multi, biz, cleanliness_rating, gu...
lgl  (3): room_shared, room_private, host_is_superhost

ℹ Use `spec()` to retrieve the full column specification for this data.
ℹ Specify the column types or set `show_col_types = FALSE` to quiet this message.
New names:
• `` -> `...1`
Rows: 4614 Columns: 20
── Column specification ────────────────────────────────────────────────────────
Delimiter: ","
chr  (1): room_type
dbl (16): ...1, realSum, person_capacity, multi, biz, cleanliness_rating, gu...
lgl  (3): room_shared, room_private, host_is_superhost

ℹ Use `spec()` to retrieve the full column specification for this data.
ℹ Specify the column types or set `show_col_types = FALSE` to quiet this message.
New names:
• `` -> `...1`
Rows: 5379 Columns: 20
── Column specification ────────────────────────────────────────────────────────
Delimiter: ","
chr  (1): room_type
dbl (16): ...1, realSum, person_capacity, multi, biz, cleanliness_rating, gu...
lgl  (3): room_shared, room_private, host_is_superhost

ℹ Use `spec()` to retrieve the full column specification for this data.
ℹ Specify the column types or set `show_col_types = FALSE` to quiet this message.
New names:
• `` -> `...1`
Rows: 4492 Columns: 20
── Column specification ────────────────────────────────────────────────────────
Delimiter: ","
chr  (1): room_type
dbl (16): ...1, realSum, person_capacity, multi, biz, cleanliness_rating, gu...
lgl  (3): room_shared, room_private, host_is_superhost

ℹ Use `spec()` to retrieve the full column specification for this data.
ℹ Specify the column types or set `show_col_types = FALSE` to quiet this message.
New names:
• `` -> `...1`
Rows: 4535 Columns: 20
── Column specification ────────────────────────────────────────────────────────
Delimiter: ","
chr  (1): room_type
dbl (16): ...1, realSum, person_capacity, multi, biz, cleanliness_rating, gu...
lgl  (3): room_shared, room_private, host_is_superhost

ℹ Use `spec()` to retrieve the full column specification for this data.
ℹ Specify the column types or set `show_col_types = FALSE` to quiet this message.

Data Wrangling and Cleaning

# Main developer: Angelyca Purewal 
# Contributors: Rapeewit Chanprakaisi

# Adding two new columns to each of the datasets 
budapest_weekdays_data <- budapest_weekdays %>% mutate(city = "Budapest", time = "Weekday")
budapest_weekends_data <- budapest_weekends %>% mutate(city = "Budapest", time = "Weekend") 
london_weekdays_data <- london_weekdays %>% mutate(city = "London", time = "Weekday")
london_weekends_data <- london_weekends %>% mutate(city = "London", time = "Weekend") 
rome_weekdays_data <- rome_weekdays %>% mutate(city = "Rome", time = "Weekday")
rome_weekends_data <- rome_weekends %>% mutate(city = "Rome", time = "Weekend") 

# Combining all the data sets into a tidy format
airbnb_data <- bind_rows(budapest_weekdays_data, budapest_weekends_data,
                         london_weekdays_data, london_weekends_data,
                         rome_weekdays_data, rome_weekends_data)

# convert boolean to binary
airbnb_data <- airbnb_data %>% mutate(room_shared = ifelse(toupper(room_shared) == "TRUE", 1, 0),
                    room_private = ifelse(toupper(room_private) == "TRUE", 1, 0),
                    host_is_superhost = ifelse(toupper(host_is_superhost) == "TRUE", 1, 0))

head(airbnb_data)

A tibble: 6 × 22

...1	realSum	room_type	room_shared	room_private	person_capacity	host_is_superhost	multi	biz	cleanliness_rating	⋯	dist	metro_dist	attr_index	attr_index_norm	rest_index	rest_index_norm	lng	lat	city	time
<dbl>	<dbl>	<chr>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	⋯	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<chr>	<chr>
0	238.9905	Entire home/apt	0	0	6	1	0	1	10	⋯	0.3593550	0.3526430	404.4047	24.116552	893.4773	67.65685	19.05074	47.50076	Budapest	Weekday
1	300.7943	Entire home/apt	0	0	6	0	0	1	9	⋯	0.9294272	0.2002355	1676.8760	100.000000	452.5397	34.26770	19.04493	47.50405	Budapest	Weekday
2	162.3819	Entire home/apt	0	0	4	1	0	0	10	⋯	2.4508403	0.2794518	163.5885	9.755551	191.9923	14.53825	19.02170	47.49882	Budapest	Weekday
3	118.4377	Entire home/apt	0	0	2	0	0	0	9	⋯	1.5594494	0.4779711	191.7198	11.433155	326.2156	24.70205	19.06301	47.51126	Budapest	Weekday
4	134.4174	Entire home/apt	0	0	4	1	1	0	10	⋯	1.1138030	0.2701016	198.6035	11.843658	635.5159	48.12322	19.06900	47.49900	Budapest	Weekday
5	127.3676	Entire home/apt	0	0	4	0	1	0	9	⋯	0.2684703	0.1669317	635.6350	37.905903	1005.6535	76.15118	19.05480	47.50094	Budapest	Weekday

# Check for missing values
cat('\nData has', sum(is.na(airbnb_data)), 'rows with missing values.')

Data has 0 rows with missing values.

Pre-selection of variables

Let's start by dropping the ...1 column that acts as an identifier as well as the room_shared and room_private columns. Since the different datasets have been merged into one, the values of the ...1 have become irrelevant, so we remove this column. Additionally, the variable room_type already categorizes the listing into private, shared, or entire home/apt, making these additional classifications unnecessary for our analysis. It is also unclear if room_private and room_shared were classifying a different aspect of the room, so we assume they represent the same information.

We'll also drop lng and lat because we haven't learned how to use spatial analysis variables.

# Main developer: Renata Lovette
# Contributors: Angelyca Purewal

# drop ID, 'room_shared' and 'room_private' columns
airbnb_data <- select(airbnb_data, -(c('...1', room_shared, room_private, lng, lat)))
head(airbnb_data)

A tibble: 6 × 17

realSum	room_type	person_capacity	host_is_superhost	multi	biz	cleanliness_rating	guest_satisfaction_overall	bedrooms	dist	metro_dist	attr_index	attr_index_norm	rest_index	rest_index_norm	city	time
<dbl>	<chr>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<chr>	<chr>
238.9905	Entire home/apt	6	1	0	1	10	99	1	0.3593550	0.3526430	404.4047	24.116552	893.4773	67.65685	Budapest	Weekday
300.7943	Entire home/apt	6	0	0	1	9	98	2	0.9294272	0.2002355	1676.8760	100.000000	452.5397	34.26770	Budapest	Weekday
162.3819	Entire home/apt	4	1	0	0	10	98	1	2.4508403	0.2794518	163.5885	9.755551	191.9923	14.53825	Budapest	Weekday
118.4377	Entire home/apt	2	0	0	0	9	92	1	1.5594494	0.4779711	191.7198	11.433155	326.2156	24.70205	Budapest	Weekday
134.4174	Entire home/apt	4	1	1	0	10	99	2	1.1138030	0.2701016	198.6035	11.843658	635.5159	48.12322	Budapest	Weekday
127.3676	Entire home/apt	4	0	1	0	9	91	2	0.2684703	0.1669317	635.6350	37.905903	1005.6535	76.15118	Budapest	Weekday

# check observations in categorical levels

# roomtype per city
rm_type_per_city_count <- airbnb_data %>%
    group_by(city) %>%
    count(room_type)

rm_type_per_city_count

A grouped_df: 9 × 3

city	room_type	n
<chr>	<chr>	<int>
Budapest	Entire home/apt	3589
Budapest	Private room	419
Budapest	Shared room	14
London	Entire home/apt	4384
London	Private room	5559
London	Shared room	50
Rome	Entire home/apt	5561
Rome	Private room	3454
Rome	Shared room	12

From the table above, the level of room_type = "Shared room" is low incomparison to the other values. If we kept this data in the dataset, there is a high chance that we would encounter problems when splitting the data into training and testing set. There is a chance that all the "Shared room" data points might fall into the testing set. Our model wouldn't recognize these points because it wouldn't have been trained on data points where room_type = "Shared room". Thus, we drop this categorical level and remove the data points accordingly.

# remove observations that include room_type = "Shared room"

airbnb_data <- airbnb_data %>% filter(room_type != "Shared room")

rm_type_per_city_count_2 <- airbnb_data %>%
    group_by(city) %>%
    count(room_type)

rm_type_per_city_count_2

A grouped_df: 6 × 3

city	room_type	n
<chr>	<chr>	<int>
Budapest	Entire home/apt	3589
Budapest	Private room	419
London	Entire home/apt	4384
London	Private room	5559
Rome	Entire home/apt	5561
Rome	Private room	3454

Finally, let's also remove rest_index and attr_index. Their normalized versions are already included in the dataset to provide a more meaningful and standardized measure, making the raw versions redundant.

airbnb_data <- airbnb_data %>% select(-rest_index, -attr_index)
head(airbnb_data)

A tibble: 6 × 15

realSum	room_type	person_capacity	host_is_superhost	multi	biz	cleanliness_rating	guest_satisfaction_overall	bedrooms	dist	metro_dist	attr_index_norm	rest_index_norm	city	time
<dbl>	<chr>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<chr>	<chr>
238.9905	Entire home/apt	6	1	0	1	10	99	1	0.3593550	0.3526430	24.116552	67.65685	Budapest	Weekday
300.7943	Entire home/apt	6	0	0	1	9	98	2	0.9294272	0.2002355	100.000000	34.26770	Budapest	Weekday
162.3819	Entire home/apt	4	1	0	0	10	98	1	2.4508403	0.2794518	9.755551	14.53825	Budapest	Weekday
118.4377	Entire home/apt	2	0	0	0	9	92	1	1.5594494	0.4779711	11.433155	24.70205	Budapest	Weekday
134.4174	Entire home/apt	4	1	1	0	10	99	2	1.1138030	0.2701016	11.843658	48.12322	Budapest	Weekday
127.3676	Entire home/apt	4	0	1	0	9	91	2	0.2684703	0.1669317	37.905903	76.15118	Budapest	Weekday

Visualization

scatter_plot <- ggplot(airbnb_data, aes(x = log(realSum), y = guest_satisfaction_overall, color = city , shape= room_type)) + 
# I have logged realSum (the price of the Airbnb listing) because this variable has extreme values therefore 
# applying the log function will compress large values while maintaining the overall trends
  geom_jitter(alpha = 0.3, width = 0.3, height = 0.3) + # jitter is used in order to spread some of the points and reduce overlapping
#I set alpha to 0.3 because lowering the alpha (opacity) makes denser areas more visible
  labs(x = "Price (Logged, EUR)", 
       y = "Guest Satisfaction", 
       color = "City", 
       shape= "Room Types",
      title = "How Price, City, and Room Type Affect Guest Satisfaction") +
      theme_minimal() +
      theme(legend.position = "right", 
      plot.title = element_text(size = 14, face = "bold"))


box_plot <- ggplot(airbnb_data, aes(x = city, y = guest_satisfaction_overall, fill = city)) +
  geom_boxplot(outlier.shape = 16, outlier.colour = "red", outlier.size = 3, alpha = 0.5) +  
  labs(x = "City", y = "Guest Satisfaction", title = "City and Guest Satisfaction") +
  theme_minimal() +
  scale_fill_manual(values = c("skyblue", "pink", "lightgreen"))

box_plot2 <- ggplot(airbnb_data, aes(x = room_type, y = guest_satisfaction_overall, fill = room_type)) +
  geom_boxplot(outlier.shape = 16, outlier.colour = "red", outlier.size = 3, alpha = 0.5) +  
  labs(x = "Room Type", y = "Guest Satisfaction", title = "Room Type and Guest Satisfaction") +
  scale_fill_manual(values = c("orange", "purple")) + 
  theme_minimal()

box_plot + box_plot2 + scatter_plot

Visualization Explanation

I have created a scatterplot to explore all 4 of my variables, and then two box plots to explore the relationship between my categorical variables and guest satisfaction. I believe these are relevant as the box plots show that the median guest satisfaction is different for each city but there is a lot of overlap and my other box plot shows that my two room types have very similar medians and overlap a lot, this tells me how much each variable plays a role in guest satisfaction. The scatterplot gives me a big picture of what is going on in my data. I had a lot of overlap so I can draw insight from that.

Methods Plan & Computational Code

Methods Plan

Method(s) of interest: Multi-covariate Linear Regression model with Lasso for variable selection

To develop a predictive model for guest satisfaction, a continuous variable, we will use multiple predictors such as the price of the listing, city (e.g., London, Rome, Budapest), location (e.g., city center distance), and listing characteristics (e.g., room type, business suitability, number of rooms). Lasso regression will be used, as it is particularly well-suited for this task by penalizing large coefficients, thus improving model interpretability and generalizability.

Before applying Lasso, we will check for multicollinearity using the Variance Inflation Factor (VIF). VIF values higher than 5 or 10 suggest multicollinearity issues that might require addressing. After checking VIF, we will split the data into a training set (80%) and a testing set (20%). The training data will be standardized and used to fit the Lasso regression model via the glmnet function. The optimal lambda (penalty term) will be selected via cross-validation. Finally, the model's performance will be evaluated using the testing set, and its prediction accuracy will be assessed by computing the Root Mean Squared Error (RMSE).

Assumptions

To use an MLR model, we need to assume the following:

• Linearity: The relationship between the dependent variable (guest_satisfaction_overall) and each independent variable should be linear.
• Independence: Observations should be independent of each other. We assume that the dataset includes a random sample of Airbnb listings.
• Homoscedasticity: The variance of residuals should remain constant across predictions.
• Normality of Residuals: Residuals should follow a normal distribution
• No extreme outliers

Potential limitations & weaknesses

• Skewed distribution: Airbnb guest satisfaction ratings tend to cluster around high values and values that are multiples of 10. Leading to a skewed distribution of the response variable. This may violate normality assumptions.

• Multicollinearity: Some predictors (e.g., location-based features) might be highly correlated, requiring ridge or lasso regression to improve model stability. Multicollinearity occurs when there is a strong association between two or more covariates. These covariates bring similar information to the model, and we have trouble isolating their effect. During the EDA stage, we've removed room_shared and room_private because they measured the same attribute. There may be other variables that will contribute to multicollinearity, so we will use VIF to check for highly correlated predictors.

Lasso will be used to mitage the previous limitation however this method adds its on limitations to the mix.

• Lasso can shrink coefficients of predictors too much, sometimes pushing important predictors to zero. This could result in an overly simplified model, losing useful information for prediction or interpretation.
• Lasso introduces bias into the model coefficients, especially for highly correlated predictors. While this helps in reducing overfitting, it can distort the true relationship between predictors and the outcome.
• Lasso assumes a linear relationship and does not automatically account for interaction terms between predictors. If there are important interactions, Lasso may not capture them unless explicitly included in the model.

# Main developer: Angelyca Purewal
# Contributors: Renata Lovette

# set seed for reproducibility
set.seed(123)

# splot the data into trainingg (80%) and testing (20%) sets
airbnb_split <-
    airbnb_data %>%
    initial_split(prop = 0.8, strata = guest_satisfaction_overall)

train_data <- training(airbnb_split)
test_data <- testing(airbnb_split)

# fit model with all predictors
mlr_model <- lm(guest_satisfaction_overall ~ ., data = train_data)  # Fit model with all predictors
summary(mlr_model)

Call:
lm(formula = guest_satisfaction_overall ~ ., data = train_data)

Residuals:
    Min      1Q  Median      3Q     Max 
-77.709  -2.143   0.604   3.163  34.427 

Coefficients:
                        Estimate Std. Error t value Pr(>|t|)    
(Intercept)           31.3788115  0.5404645  58.059  < 2e-16 ***
realSum                0.0002273  0.0001553   1.464 0.143194    
room_typePrivate room  0.5235988  0.1232109   4.250 2.15e-05 ***
person_capacity       -0.0099721  0.0519793  -0.192 0.847864    
host_is_superhost      1.5669630  0.1118002  14.016  < 2e-16 ***
multi                 -0.7798785  0.1183055  -6.592 4.46e-11 ***
biz                   -2.5321455  0.1203530 -21.039  < 2e-16 ***
cleanliness_rating     6.6375584  0.0491870 134.945  < 2e-16 ***
bedrooms               0.3665806  0.0992950   3.692 0.000223 ***
dist                  -0.0144839  0.0377551  -0.384 0.701258    
metro_dist             0.2896782  0.0636307   4.552 5.34e-06 ***
attr_index_norm       -0.0032485  0.0081587  -0.398 0.690510    
rest_index_norm        0.0095233  0.0050614   1.882 0.059911 .  
cityLondon            -1.6897013  0.2488910  -6.789 1.16e-11 ***
cityRome              -1.8308295  0.1436378 -12.746  < 2e-16 ***
timeWeekend           -0.0326030  0.0927098  -0.352 0.725092    
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 6.265 on 18356 degrees of freedom
Multiple R-squared:  0.5678,	Adjusted R-squared:  0.5674 
F-statistic:  1607 on 15 and 18356 DF,  p-value: < 2.2e-16

# use VIF to conduct an initial check for multicolinearity – values higher than 5 or 10 suggest multicollinearity might be problematic.
vif(mlr_model) %>% round(3)

A matrix: 14 × 3 of type dbl

	GVIF	Df	GVIF^(1/(2*Df))
realSum	1.285	1	1.133
room_type	1.724	1	1.313
person_capacity	2.160	1	1.470
host_is_superhost	1.134	1	1.065
multi	1.433	1	1.197
biz	1.561	1	1.249
cleanliness_rating	1.120	1	1.058
bedrooms	1.592	1	1.262
dist	4.495	1	2.120
metro_dist	1.881	1	1.371
attr_index_norm	3.391	1	1.842
rest_index_norm	2.818	1	1.679
city	5.315	2	1.518
time	1.005	1	1.003

The table above shows low to moderate multicolinearity so we will use Lasso for further variable selection.

X_train <- model.matrix(guest_satisfaction_overall ~ ., data = train_data)[, -1]  # Remove intercept
y_train <- train_data$guest_satisfaction_overall
X_test <- model.matrix(guest_satisfaction_overall ~ ., data = test_data)[, -1]
y_test <- test_data$guest_satisfaction_overall

# Standardize predictors
X_train_scaled <- scale(X_train)
X_test_scaled <- scale(X_test, center = attr(X_train_scaled, "scaled:center"), scale = attr(X_train_scaled, "scaled:scale"))

# Perform Lasso Regression with cross-validation
cv_lasso <- cv.glmnet(X_train_scaled, y_train, alpha = 1)  # Lasso (alpha = 1)

# Get the best lambda
best_lambda_lasso <- cv_lasso$lambda.min

# Train the final Lasso model using best lambda
lasso_model <- glmnet(X_train_scaled, y_train, alpha = 1, lambda = best_lambda_lasso)

# see variables selected by Lasso
lasso_coefs <- coef(lasso_model, s = best_lambda_lasso)
lasso_coefs

16 x 1 sparse Matrix of class "dgCMatrix"
                                s1
(Intercept)           92.317276290
realSum                0.049296255
room_typePrivate room  0.226381498
person_capacity        .          
host_is_superhost      0.682992564
multi                 -0.335509143
biz                   -1.191989922
cleanliness_rating     6.595315479
bedrooms               0.195707769
dist                   .          
metro_dist             0.261768019
attr_index_norm        .          
rest_index_norm        0.127797759
cityLondon            -0.818598691
cityRome              -0.848788478
timeWeekend           -0.003277905

Lasso has shrunk person_capacity, dist and attr_index_norm variables to 0. We will now compare the performance of the Lasso model to the full MLR model.

test_pred_full <- predict(mlr_model, newdata = test_data)

# compute RMSE of the full predictive model
airbnb_test_RMSEs <- tibble(
  Model = "OLS Full Regression",
  RMSE = mltools::rmse(
    preds = test_pred_full,
    actuals = test_data$guest_satisfaction_overall
  )
)

# compute RMSE of lasso predictive model
test_pred_lasso <- predict(lasso_model, s = best_lambda_lasso, newx = X_test_scaled)

airbnb_test_RMSEs <- rbind(
  airbnb_test_RMSEs,
  tibble(
    Model = "Lasso with minimum MSE",
    RMSE = mltools::rmse(
       pred = test_pred_lasso,
       actuals = y_test)
  )
)
airbnb_test_RMSEs

A tibble: 2 × 2

Model	RMSE
<chr>	<dbl>
OLS Full Regression	6.127537
Lasso with minimum MSE	6.127711

Analysing RMSE values

The RMSE values of the full MLR model and the Lasso model are nearly identical with a difference of 0.000174, which is practically negligible. The Lasso model excluded 3 coefficients. Due to the very minor difference, it suggests that our data doesn't suffer much from multicollinearity.

Visualizing Predictions

# Main developer: Angelyca Purewal

# Create a data frame with actual and predicted values
prediction_data <- data.frame(Actual = test_data$guest_satisfaction_overall, 
                              s1 = lasso_preds)

# Scatter plot: Predicted vs Actual
predicted_actual_plot <- ggplot(prediction_data, aes(x = Actual, y = s1)) +
  geom_jitter(alpha = 0.3, width = 0.3, height = 0.3) + # jitter is used in order to spread some of the points and reduce overlapping
  geom_point(color = "blue", alpha = 0.5) +  # Points for Predicted vs Actual values
  geom_abline(slope = 1, intercept = 0, color = "red", linetype = "dashed") +  # Line of perfect prediction
  labs(title = "Predicted vs Actual Guest Satisfaction", 
       x = "Actual Guest Satisfaction", 
       y = "Predicted Guest Satisfaction") +
  theme_minimal() +
  theme(legend.position = "right", 
  plot.title = element_text(size = 14, face = "bold"))

predicted_actual_plot

Visualization Interpretation

If this model was perfect, all points would fall on the red line, instead, we have a spread around the line, indicating prediction error. This model predicts high guest satisfaction fairly well, because most of our training data consisted of higher ratings, which is also why this model struggles to predict low satisfaction scores, often overestimating them. The predictions are also compressed toward the average, a common effect of Lasso regression shrinking extreme values. To improve, we would need to address the imbalance in low-rated listings.

Discussion

This analysis demonstrated that guest satisfaction on Airbnb can be partially predicted using listing characteristics and location data. While Lasso regression provided a good balance between bias and variance, several challenges and limitations affected the model’s accuracy. Price, cleanliness, and location-based variables were among the most predictive features. Guest satisfaction tends to be biased upward, with many listings having high ratings, complicating the prediction of lower scores.

The model provided a reasonable fit for most observations, especially those with satisfaction scores clustered around the average. Lasso regression helped to reduce overfitting and offered interpretability by automatically selecting relevant variables. The low RMSE on the test set suggests that our model has moderate predictive power for new listings.

Our results emphasise that satisfaction is multifactorial, driven not only by price but also by location, cleanliness, and room features. The importance of cleanliness and amenities in increasing guest loyalty and satisfaction.(Xu and Gursoy, 2020) For hosts, improving cleanliness ratings and carefully setting price levels may improve the guest experience. For Airbnb as a platform, refining satisfaction metrics like collecting more detailed feedback or using a broader rating scale may support better model training and platform trust.

Limitations

The guest satisfaction variable has a left-skewed distribution, with most values concentrated near 100. This limits the model’s ability to distinguish well between average and excellent listings and to identify factors contributing to dissatisfaction. And a few listings received low satisfaction scores, which limits the model’s ability to learn patterns from negative feedback. It may cause an imbalance in the response variable and then introduce bias in model evaluation.

To improve model performance, we suggest implementing ensemble models like Random Forest or Gradient Boosted Trees, which can better capture nonlinear relationships and interaction effects.

Future Research

Future research could incorporate temporal features such as booking dates, seasons, and day-of-week effects to analyze how guest satisfaction varies over time. For example, satisfaction may dip during peak tourist seasons due to higher prices or stretched host resources, while winter off-season listings might show higher satisfaction due to quieter environments and better host availability. Including time-based variables could uncover seasonal patterns and improve model accuracy by capturing temporal dynamics often overlooked in static models.

References

Airbnb prices in European cities. (2024, March 10). Kaggle. https://www.kaggle.com/datasets/thedevastator/airbnb-prices-in-european-cities

Gyódi, K., & Nawaro, Ł. (2021). Determinants of Airbnb prices in European cities: A spatial econometrics approach (Supplementary Material) [Dataset]. In Zenodo (CERN European Organization for Nuclear Research). https://doi.org/10.5281/zenodo.4437019

Rezazadeh Kalehbasti, P., Nikolenko, L., & Rezaei, H. (2021). Airbnb price prediction using machine learning and sentiment analysis. In E. Weippl, A. Holzinger, A. M. Tjoa & P. Kieseberg (Eds.), Machine learning and knowledge extraction (pp. 173-184). Springer International Publishing AG. https://doi.org/10.1007/978-3-030-84060-0_11

Xu, X., & Gursoy, D. (2020). Exploring the relationship between servicescape, place attachment, and intention to recommend accommodations marketed through sharing economy platforms. Journal of Travel & Tourism Marketing, 37(4), 429–446. https://doi.org/10.1080/10548408.2020.1784365